Re: [Pocl-devel] async_work_group_copy

[Erik S. <esc...@pe...>]

2012/1/22 Pekka Jääskeläinen <pek...@tu...>:
> Hi Erik,
>
>
> On 01/20/2012 07:05 PM, Erik Schnetter wrote:
>>
>> 3-element vectors can't be accessed directly in memory, because they
>> need to be aligned the same was as 4-vectors. One needs to use
>> vload/vstore for this, which doesn't require alignment.
>>
>> I assume that 3-element vectors would be stored packed in global
>> memory. I don't know whether this is what async_copy actually
>> expects... However, un-packed storage can be handled by passing a
>> respective 4-element vector as gentype.
>
>
> I'm not sure if I got this. The current implementation of
> the async copy is just a for loop that copies elements of
> the arrays.
>
> The elements in the copy loop are handled as elements of
> the actual gentype, e.g. float3. As far as I understood
> this means the compiler should take care of the correct alignment of
> the vectors in memory because it's stated with the alignment
> attribute (to be the same as float4) both in the host side and the
> kernel side.
>
> Did you check from the generated assembly that the copy loop does
> not produce a correct alignment when copying buffers with 3-element
> vectors?
>
> "The vload3 and vstore3 built-in functions can be used to read and
> write, respectively, 3-component vector data types from an array of
> packed scalar data type."
>
> From cl_platform.h:
>
> /* cl_float3 is identical in size, alignment and behavior to cl_float4. See
> section 6.1.5. */
> typedef  cl_float4  cl_float3;
>
>
> So, my understanding of this is that it should work because both the
> host and the device code are in the same understanding of the alignment
> of the 3-element vectors. I.e. it's stored 4-aligned everywhere if
> you handle buffers that contain 3-element vectors and thus copy loop
> should be generated to code that adheres to this.
>
> If the input was packed (basically a float buffer) and one would like
> to copy it to a buffer of 3-wide vectors, then vload3 and vstore3 should
> be used.
>
> Please correct me if I understood this wrongly.

Yes, this is correct. I was assuming that packed buffers were more
common than unpacked ones (otherwise, why would one use float3?), and
that async_copy would need to handle this. I now realise that there is
no basis for this assumption of mine. Indeed there is a footnote in
the standard stating that float3 should be copied in the same way as
float4.

Your implementation is correct.

-erik

-- 
Erik Schnetter <esc...@pe...>
http://www.cct.lsu.edu/~eschnett/
AIM: eschnett247, Skype: eschnett, Google Talk: sch...@gm...